Nvidia offered an early look at its next rack-scale AI system, Vera Rubin, which the company says delivers 10 times the performance per watt of its Grace Blackwell predecessor. The modular, fully liquid-cooled platform pairs 72 Rubin GPUs with 36 Vera CPUs and packs roughly 1,300 chips and 1.3 million total components per rack. While Rubin will draw about twice the power of Blackwell, Nvidia argues its efficiency gains will lower the cost of inference and training on a tokens-per-power basis—an increasingly critical metric as data-center energy use climbs. The system is expected to ship in the second half of 2026, with estimated rack pricing 25% above Blackwell at roughly $3.5 million to $4 million. Nvidia says it has lined up major buyers, including Meta, OpenAI, Anthropic, Amazon, Google and Microsoft, and is coordinating a global supply chain of more than 80 vendors amid tight memory markets. The launch comes as competition intensifies from AMD’s forthcoming Helios rack-scale system and custom silicon from Broadcom and Google, with hyperscalers seeking a credible second source.
Related articles:
– Cloud TPU
– AMD Instinct
– TSMC Arizona
– High Bandwidth Memory
